any · renderer · library

3Blue1Brown and manim (and agents writing manim)

Python engine for programmatic math animation, now a target for agents generating explainer videos.

inspireevidence: strongby Grant Sanderson; Manim Communitygithub.com/ManimCommunity/manim ↗

3Blue1Brown and manim (and agents that write manim)

  • Maker: Grant Sanderson (3Blue1Brown); the community fork by the Manim Community
  • URL: https://github.com/3b1b/manim and https://github.com/ManimCommunity/manim
  • Status (10 October 2026): 3b1b/manim has 94.8k stars (MIT, last push 9 September 2026); ManimCommunity/manim has 41.4k stars (MIT, v0.22.0 on 8 October 2026). Both active. Research systems that generate manim with LLMs include TheoremExplainAgent (TIGER-AI-Lab, February 2025, 1.5k stars), Manimator (ICML 2025) and LLM2Manim (2026), plus community Claude Code skills.

What it is

manim is the Python animation engine Sanderson wrote for the 3Blue1Brown videos: you describe mathematical objects and transformations in code and it renders a video. The community fork is the documented, stable one most people use.

The problem it’s solving

Mathematical ideas are about change and structure, which static diagrams and text show badly. Precise, programmatic animation makes the transformation itself visible.

Its path / bet

Code in, video out. The explanation is authored, scripted and narrated; the viewer watches, they don’t interact. With LLMs, the bet shifts: an agent writes the scene code. TheoremExplainAgent reports a 93.8% success rate generating long theorem videos with an o3-mini agent, while noting that most videos had minor layout problems; LLM2Manim keeps a human review step for that reason.

How it works (concretely)

Python scenes made of mobjects (shapes, LaTeX, graphs) and animations (Transform, FadeIn), rendered with Cairo or OpenGL to video. Agent pipelines plan a storyboard, write scene code, render, look at the result and repair.

Strengths

  • The most successful visual-maths explanation tool there is, by audience.
  • Code is a precise, checkable description, which suits agents.

Weaknesses / limits

  • Output is a video: not interactive, not bound to data, not readable back by an agent except by watching frames.
  • Rendering is slow (seconds to minutes per scene).
  • Generated videos have layout errors often enough that research pipelines keep humans in the loop.

Relation to fictty

Inspire. It shows what a good visual explanation looks like and that agents can now author one. It sits at the opposite corner from fictty: authored code, rendered offline, watched passively.

Could fictty adopt it instead of building?

No. Different medium. A fictty screen could show a rendered manim clip through pixels on a kitty terminal, but that’s a curiosity.

What fictty should take from it

  • Animate the data, not the pictures (already fictty guidance) is what manim does well: a transform between two states. A fictty transition between two values of a bound node, shown as steps, is the terminal version.
  • Layout errors are the failure mode of generated visuals. A component library with layout built in, as fictty has, avoids the class of bug that dominates manim generation.

Sources